Diffusion models in structural biology and protein science

How it is used

Structural biology works out the shapes that proteins fold into, because shape largely decides what a protein does and how drugs might bind to it. Measuring those shapes by experiment is slow, and many proteins resist it altogether. A diffusion model offers another route. It is trained by taking known structures, blurring them step by step into random noise, and learning to undo each step. Run in reverse from pure noise, it draws a plausible structure, and it can be steered by a sequence, a binding partner or a fixed scaffold. The newer structure-prediction systems use this denoising step to place atoms, and related models design sequences and backbones outright.

In practice the models are usually run off the shelf rather than retrained. Researchers feed in sequences and ask for full-length models of proteins that have never been solved experimentally, including complexes of two chains, then compare those models with each other or with crystal structures using standard alignment and docking scores. The same systems are turned on harder targets, such as a small peptide binding into a receptor, where the answer is often sampled many times with different random starting points to see how stable it is. Confidence values the model reports about its own prediction are then checked against how far the prediction actually sits from the experimental structure.

Why researchers use it

The appeal is coverage. Most sequences in the databases have no solved structure, and a model that produces a coordinate file in hours stands in for months of crystallography or cryo-electron microscopy that may never succeed. Once a structure exists, however approximate, the ordinary tools of the field become available: superposing two proteins, measuring how far equivalent atoms sit apart, finding which residue in a yeast protein matches a position mutated in a human cancer.

Generation, rather than prediction, is the second draw. Because the model samples instead of returning one fixed answer, it can be asked for many candidates and run inside a loop that scores each one and pushes the sequence towards a wanted fold. That turns protein design into a search, with the model supplying the proposals and a separate scoring step supplying the judgement. Repeated sampling also serves as a rough check on the model itself, since predictions that shift with the random seed are ones to distrust.

In the record

Limits

A generated structure is a guess shaped by what the model saw in training. Proteins and complexes unlike anything in the databases are predicted worst, and short or flexible pieces, such as a peptide binding a receptor, are shakier than a compact folded domain. The confidence scores these models report do not always track the real error. Sampling the same input repeatedly can give answers that differ, and nothing in the method says which is right. A model also shows one static shape, while real proteins move, and a designed sequence remains a proposal until it is made and tested.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

• Diffusion model: the idea, across every field
• AI in structural biology and protein science

Every article